CS 176

Project 1

Papers:
Wei and Levoy Fast Texture Synthesis using Tree-Structured Vector Quantization
Ashikhmin Synthesizing Natural Textures

Data:
Sample (ie input) Textures are 128x128 jpeg images
(the first six input images are from the project website. the seventh image is a resized, cropped picture I took of the Chester pool during heavy rains on Monday using my Canon s45)
Output Textures are 512x512 png images

Code:
I used Corona to input and output images. The code was compiled using Visual C++ 6.0 running under Windows XP Pro SP2 on a P4 2.26Ghz Northwood with 512MB DDR333 RAM. The code is downloadable here. The runtimes were output to file. The runtimes for exhaustive search are here and the runtimes for candidate pixel optimization are here. The runtimes increase over time, since the runs are made back-to-back and I was too lazy to clear the memory and whatnot.

Runtimes are approximate (in seconds):
Exhaustive Search:
5x5 Window: ~180
7x7 Window: ~250
9x9 Window: ~500
Candidate Pixel Optimization
5x5 Window: ~1.0
7x7 Window: ~1.4
9x9 Window: ~2.5

Analysis:
Comparing runtimes to that quoted by Wei and Levoy: 360 sec for 200x200 texture using a 5x5 window (195Mhz R10000)
180 sec for 512x512 texture using a 5x5 window (2266Mhz P4 Northwood)
We expect the 512x512 texture to take (512x512)/(200x200) = 6.5536 longer to run on the same processor.
This means that my runtimes are ~13 times faster, which is pretty good when reviewing other benchmarks comparing p4's and the R10000, and also considering p4's long pipeline.
I did generate texture windows for each pixel of the texture and stored that in memory to compare against the neighborhoods in the sample. This was so that the memory to be accessed was a significantly smaller array (5x3 - 9x5) than the whole texture array (512x512). Presumable, optimizations would be able to place this in the CPU L2 cache, which is 512k, instead of RAM. The texture array would not be able to fit in the cache. I have yet to see what would happen if I ran my code on a processor that has 2mb of cache, such as a G5, P4EE, or Banias (pentium m).

Results:
Click on thumbnails for full-size versions.
Exhaustive Search Candidate Pixel Optimization
Input Texture Generated Texture Generated Texture Generated Texture Generated Texture Generated Texture Generated Texture
128x128 512x512 512x512 512x512 512x512 512x512 512x512
Window Size: 5x5 Window Size: 7x7 Window Size: 9x9 Window Size: 5x5 Window Size: 7x7 Window Size: 9x9
Green Pattern Sample Green Pattern generated using Exhaustive Search with window size 5x5 Green Pattern generated using Exhaustive Search with window size 7x7 Green Pattern generated using Exhaustive Search with window size 9x9 Green Pattern generated using Candidate Pixel Optimization with window size 5x5 Green Pattern generated using Candidate Pixel Optimization with window size 7x7 Green Pattern generated using Candidate Pixel Optimization with window size 9x9
Grid Sample Grid generated using Exhaustive Search with window size 5x5 Grid generated using Exhaustive Search with window size 7x7 Grid generated using Exhaustive Search with window size 9x9 Grid generated using Candidate Pixel Optimization with window size 5x5 Grid generated using Candidate Pixel Optimization with window size 7x7 Grid generated using Candidate Pixel Optimization with window size 9x9
Orange Flowers Sample Orange Flowers generated using Exhaustive Search with window size 5x5 Orange Flowers generated using Exhaustive Search with window size 7x7 Orange Flowers generated using Exhaustive Search with window size 9x9 Orange Flowers generated using Candidate Pixel Optimization with window size 5x5 Orange Flowers generated using Candidate Pixel Optimization with window size 7x7 Orange Flowers generated using Candidate Pixel Optimization with window size 9x9
Peppers Sample Peppers generated using Exhaustive Search with window size 5x5 Peppers generated using Exhaustive Search with window size 7x7 Peppers generated using Exhaustive Search with window size 9x9 Peppers generated using Candidate Pixel Optimization with window size 5x5 Peppers generated using Candidate Pixel Optimization with window size 7x7 Peppers generated using Candidate Pixel Optimization with window size 9x9
Purple Flowers Sample Purple Flowers generated using Exhaustive Search with window size 5x5 Purple Flowers generated using Exhaustive Search with window size 7x7 Purple Flowers generated using Exhaustive Search with window size 9x9 Purple Flowers generated using Candidate Pixel Optimization with window size 5x5 Purple Flowers generated using Candidate Pixel Optimization with window size 7x7 Purple Flowers generated using Candidate Pixel Optimization with window size 9x9
Stones Sample Stones generated using Exhaustive Search with window size 5x5 Stones generated using Exhaustive Search with window size 7x7 Stones generated using Exhaustive Search with window size 9x9 Stones generated using Candidate Pixel Optimization with window size 5x5 Stones generated using Candidate Pixel Optimization with window size 7x7 Stones generated using Candidate Pixel Optimization with window size 9x9
Water Sample Water generated using Exhaustive Search with window size 5x5 Water generated using Exhaustive Search with window size 7x7 Water generated using Exhaustive Search with window size 9x9 Water generated using Candidate Pixel Optimization with window size 5x5 Water generated using Candidate Pixel Optimization with window size 7x7 Water generated using Candidate Pixel Optimization with window size 9x9

Valid XHTML 1.0!